This chapter defines artificial intelligence (AI) from the perspective of the AI-900 exam, focusing on its core concepts, workloads, and principles. Understanding what AI is and what it is not is foundational for all other topics in the exam. Approximately 10-15% of AI-900 questions directly test your grasp of AI definitions, common workloads, and the distinction between AI, machine learning, and traditional programming. This chapter will equip you with the precise terminology and mental models to answer these questions correctly and avoid common traps.
Jump to a section
Imagine a master chef who has never cooked a dish before but has access to a library of thousands of recipes. Each recipe is a set of instructions (an algorithm) that tells the chef exactly what ingredients to use, in what order, and for how long to cook. The chef can follow a recipe perfectly every time without understanding why it works. This is like traditional programming: the computer follows explicit instructions. Now imagine the chef is asked to create a new dish that tastes like a particular cuisine, but no recipe exists. The chef studies hundreds of existing recipes, identifies patterns (e.g., certain spices appear together, cooking times correlate with ingredient types), and then experiments by combining ingredients in novel ways, tasting and adjusting. Over many trials, the chef learns which combinations work. This is machine learning: the AI learns patterns from data rather than following hard-coded rules. The chef's ability to recognize a dish by tasting it (classification) or to predict how long to cook a new ingredient (regression) mirrors AI workloads like computer vision or predictive analytics. The key mechanistic parallel is that the chef does not memorize every possible dish; instead, they internalize statistical relationships from examples, enabling generalization to new inputs.
What is Artificial Intelligence?
Artificial Intelligence (AI) is the simulation of human intelligence processes by machines, especially computer systems. These processes include learning (acquiring information and rules for using it), reasoning (using rules to reach approximate or definite conclusions), and self-correction. For the AI-900 exam, you must understand that AI is not a single technology but a field encompassing many subfields, including machine learning, natural language processing, computer vision, and speech recognition.
Why AI Exists
Traditional programming requires explicit instructions for every possible scenario. This approach fails for tasks that are too complex to codify manually, such as recognizing a face in a photo, understanding spoken language, or predicting stock prices. AI systems learn patterns from data, enabling them to handle variability and make decisions in novel situations. The key insight is that AI excels at tasks where the rules are not known or are too difficult to express, but examples are abundant.
How AI Works Internally
At its core, AI relies on algorithms that process data to produce outputs. The most common approach in modern AI is machine learning (ML). In ML, a model is trained using a dataset of examples. The model learns statistical patterns that map inputs to outputs. For instance, a spam filter learns that emails containing certain words (e.g., "free", "win") are likely spam. The training process involves:
Data Collection: Gathering a large, representative dataset.
Feature Extraction: Identifying relevant attributes (e.g., pixel values in an image).
Model Selection: Choosing an algorithm (e.g., decision tree, neural network).
Training: The algorithm adjusts its internal parameters to minimize prediction error on the training data.
Evaluation: Testing the model on unseen data to ensure it generalizes.
Once trained, the model can infer outputs for new inputs. This is called inference.
Key Components of AI Systems
Data: The fuel for AI. Quality, quantity, and relevance of data directly impact performance.
Algorithms: The mathematical procedures that learn from data. Examples include linear regression, support vector machines, and deep neural networks.
Models: The output of training—a representation of the learned patterns. Models are deployed to make predictions.
Compute Resources: Training large models requires significant computational power, often using GPUs or TPUs.
AI vs. Machine Learning vs. Deep Learning
AI: Broad field aiming to create intelligent machines.
Machine Learning: Subset of AI where machines learn from data without being explicitly programmed.
Deep Learning: Subset of ML using neural networks with many layers (deep neural networks) to model complex patterns.
On the exam, remember that all deep learning is machine learning, and all machine learning is AI, but not vice versa.
Common AI Workloads on Azure
Microsoft Azure offers several AI services aligned with common workloads:
Computer Vision: Analyzing images and videos. Azure Computer Vision can extract text, describe scenes, detect objects.
Natural Language Processing (NLP): Understanding text and speech. Azure Language Service provides sentiment analysis, key phrase extraction, language detection.
Speech Recognition and Synthesis: Converting speech to text and vice versa. Azure Speech Service includes speech-to-text, text-to-speech, and translation.
Decision Support: Systems that recommend actions. Azure Personalizer uses reinforcement learning to personalize content.
Anomaly Detection: Identifying unusual patterns. Azure Anomaly Detector flags outliers in time series data.
Responsible AI Principles
Microsoft promotes six principles for responsible AI:
Fairness: AI systems should treat all people fairly.
Reliability and Safety: Systems should operate reliably and safely.
Privacy and Security: Must protect user data.
Inclusiveness: Should empower everyone and engage diverse perspectives.
Transparency: Users should understand how decisions are made.
Accountability: Humans should be accountable for AI systems.
These principles are tested on the exam—know them by name and definition.
The AI Continuum: From Narrow to General AI
Narrow AI (Weak AI): Designed to perform a specific task, like playing chess or recommending movies. All current AI is narrow.
General AI (Strong AI): Would have human-like cognitive abilities across any domain. This does not exist yet.
The exam may ask you to identify examples of narrow AI (e.g., a chatbot) versus general AI (which is hypothetical).
How AI Interacts with Related Technologies
Big Data: AI often requires large datasets for training. Big data technologies like Azure Data Lake Store provide storage.
Cloud Computing: Cloud platforms offer scalable compute and pre-built AI services. Azure AI services are consumed via REST APIs.
IoT: AI can analyze data from IoT devices for predictive maintenance or anomaly detection.
RPA (Robotic Process Automation): AI can enhance RPA by adding decision-making capabilities (e.g., reading invoices using OCR).
Common Misunderstandings
AI is not magic: It is based on mathematical optimization. Models are only as good as their data.
AI does not think like humans: It processes data statistically, not with consciousness.
AI can be biased: If training data contains biases, the model will learn them.
Exam-Specific Details
AI-900 Objective 1.1: "Identify common AI workloads." You must be able to match scenarios to appropriate Azure AI services.
Key terms: Inference, training, features, labels, supervised vs. unsupervised learning, regression, classification.
Default values: Not applicable for this conceptual topic, but you should know that Azure Cognitive Services (now Azure AI Services) offer free tiers with limits (e.g., 5,000 transactions per month for some services).
Configuration and Verification
While this chapter is conceptual, you can experiment with Azure AI services via the Azure portal or REST APIs. For example, to test Computer Vision:
curl -H "Ocp-Apim-Subscription-Key: <your-key>" -H "Content-Type: application/json" "https://<region>.api.cognitive.microsoft.com/vision/v3.2/analyze?visualFeatures=Description,Tags" -d '{"url":"https://example.com/image.jpg"}'This returns a JSON object with descriptions and tags.
Summary
AI is a broad field focused on creating systems that can perform tasks requiring human intelligence. The AI-900 exam tests your understanding of AI workloads, the relationship between AI, ML, and deep learning, and responsible AI principles. Focus on recognizing which Azure service fits a given scenario and the terminology of model training and inference.
Identify the Business Problem
Before building an AI solution, clearly define the problem. Ask: Is it a classification task (e.g., is this email spam?), regression (predict a numeric value), or clustering (group similar customers)? This determines the type of AI workload. For the exam, map scenarios to Azure services: a problem involving image analysis suggests Computer Vision; text analysis suggests Language Service. Document the success criteria, data availability, and constraints like latency or cost.
Collect and Prepare Data
Data is the foundation. Gather a dataset that represents the real-world distribution. For supervised learning, ensure each example has a label. Clean data by handling missing values, removing duplicates, and normalizing features. For image data, resize and augment (e.g., rotate, flip) to increase diversity. In Azure, you can use Azure Machine Learning data stores and datasets. The quality of data directly impacts model accuracy—garbage in, garbage out.
Choose an Algorithm and Train
Select an algorithm based on the problem type and data characteristics. For structured data, decision trees or logistic regression work well. For unstructured data (images, text), deep learning models like convolutional neural networks (CNNs) are common. Split data into training (70-80%), validation (10-15%), and test (10-15%) sets. Train the model by feeding training data and adjusting parameters to minimize loss. Use Azure Machine Learning's automated ML to try multiple algorithms.
Evaluate and Tune the Model
Evaluate the model on the validation set using metrics like accuracy, precision, recall, F1 score for classification, or mean absolute error (MAE) for regression. Tune hyperparameters (e.g., learning rate, number of layers) to improve performance. Avoid overfitting—where the model memorizes training data but fails on new data—by using regularization or cross-validation. In Azure, you can use hyperdrive for hyperparameter tuning.
Deploy and Monitor
Deploy the trained model as a web service for inference. In Azure, you can deploy to Azure Kubernetes Service (AKS), Azure Container Instances (ACI), or as a real-time endpoint. Monitor performance using Azure Monitor and Application Insights. Retrain periodically with new data to prevent concept drift. For the exam, know that deployment options include real-time endpoints and batch endpoints.
Enterprise Scenario 1: Customer Service Chatbot
A large telecom company deploys an AI-powered chatbot to handle customer inquiries. Problem: High call volume leads to long wait times and customer dissatisfaction. Solution: Use Azure Bot Service integrated with Language Service for natural language understanding. The bot is trained on historical chat logs to classify intents (e.g., 'billing question', 'technical support') and extract entities (e.g., account number, date). The bot handles common queries automatically, escalating complex issues to human agents. Configuration: The bot uses a QnA Maker knowledge base for FAQs and Language Understanding (LUIS) for intent classification. Scale: Handles 10,000+ conversations daily with 99.9% uptime. Performance: Latency under 2 seconds per response. Common misconfiguration: Inadequate training data leads to poor intent recognition, causing the bot to misunderstand users. Regular retraining with new conversation logs is essential to maintain accuracy.
Enterprise Scenario 2: Predictive Maintenance in Manufacturing
A manufacturing plant uses AI to predict equipment failures before they occur. Problem: Unexpected machine breakdowns cause costly production downtime. Solution: IoT sensors collect vibration, temperature, and pressure data from machinery. Azure Anomaly Detector analyzes time series data to detect unusual patterns indicating impending failure. Additionally, a regression model predicts remaining useful life (RUL) of components. Configuration: Data is ingested via Azure IoT Hub into Azure Stream Analytics for real-time processing. The anomaly detection model is trained on historical failure data. Scale: Monitors 500+ machines, processing 1GB of sensor data per hour. Performance: Alerts generated within 1 minute of anomaly detection. Common misconfiguration: Setting anomaly sensitivity too high causes false alarms; too low misses real failures. Tuning the sensitivity parameter (default: 95% confidence) is critical.
Enterprise Scenario 3: Document Processing with OCR
An insurance company automates claims processing using AI. Problem: Manual data entry from paper forms and PDFs is slow and error-prone. Solution: Azure Form Recognizer extracts key-value pairs (e.g., policy number, date of loss) from documents using optical character recognition (OCR) and custom models. The extracted data is validated and fed into the claims system. Configuration: Prebuilt models for invoices and receipts are used initially; custom models are trained on specific form layouts. Scale: Processes 50,000 documents per month. Performance: Extraction accuracy >95% after training. Common misconfiguration: Poor document quality (low resolution, skewed text) reduces OCR accuracy. Preprocessing steps like deskewing and contrast adjustment improve results.
AI-900 Objective 1.1: Identify Common AI Workloads
This objective tests your ability to match business scenarios to the appropriate Azure AI service. The exam presents a brief scenario (e.g., 'A company wants to extract key phrases from customer reviews') and asks which service to use. Key services to know:
Computer Vision: Image analysis, OCR, object detection.
Language Service: Sentiment analysis, key phrase extraction, language detection.
Speech Service: Speech-to-text, text-to-speech, speech translation.
Decision Services: Anomaly Detector, Personalizer, Content Moderator.
Common Wrong Answers and Why
Confusing Computer Vision with Language Service: A scenario describing 'extracting text from images' requires Computer Vision (OCR), not Language Service. Candidates often pick Language Service because they think of text, but the input is an image.
Choosing Cognitive Services instead of Azure AI Services: Microsoft recently rebranded 'Cognitive Services' to 'Azure AI Services'. The exam may use either term, but know they are the same. Wrong answer: 'It's a different offering.'
Misidentifying Anomaly Detector: Candidates might think Anomaly Detector is for fraud detection in transactions (yes), but also for any time series data. A scenario about 'detecting unusual patterns in server metrics' is also Anomaly Detector.
Specific Numbers and Terms
Free tier limits: Many Azure AI services offer a free tier with 5,000 transactions per month (e.g., Computer Vision, Language Service). This is not directly tested but helpful for context.
Key terms: Inference, training, features, labels, supervised vs. unsupervised learning.
Responsible AI principles: Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, Accountability. Expect a question asking you to identify which principle is violated in a scenario.
Edge Cases and Exceptions
Custom Vision vs. Computer Vision: Custom Vision allows you to train custom image classification/object detection models. The exam may ask: 'Which service should you use to identify specific types of defects in manufactured parts?' Answer: Custom Vision, not Computer Vision (which is prebuilt).
Speech Translation: This is part of Speech Service, not Translator Service. Translator is for text translation.
Personalizer: Used for content recommendations (e.g., which article to show a user). It uses reinforcement learning, not supervised learning.
How to Eliminate Wrong Answers
Identify the input type: Image? Use Computer Vision. Text? Use Language Service. Audio? Use Speech Service.
Identify the output needed: Description? Computer Vision. Sentiment? Language Service. Translation? Speech or Translator.
If custom training is needed: Custom Vision, Custom Speech, Custom Translator, or Custom Text (Language Service).
If the scenario mentions 'real-time' or 'batch': Both are possible; the exam usually doesn't differentiate unless specific.
By systematically mapping inputs and outputs to services, you can eliminate distractors.
AI is a broad field; ML is a subset; deep learning is a subset of ML.
Common AI workloads on Azure: Computer Vision, Language, Speech, Decision, Anomaly Detection.
Responsible AI principles: Fairness, Reliability, Privacy, Inclusiveness, Transparency, Accountability.
Narrow AI (weak) performs specific tasks; General AI (strong) does not exist yet.
Training is the process of teaching a model using data; inference is using the model to make predictions.
Azure AI Services (formerly Cognitive Services) provide pre-built AI capabilities via REST APIs.
Custom Vision allows training custom image models; Language Service provides pre-built NLP.
Bias in AI comes from biased training data; it is not automatically removed.
These come up on the exam all the time. Here's how to tell them apart.
Traditional Programming
Rules are explicitly coded by programmers.
Output is deterministic based on input.
Cannot handle unseen scenarios without new code.
Best for well-defined, stable problems.
Example: A tax calculator follows fixed tax rules.
Machine Learning
Model learns rules from data automatically.
Output is probabilistic; may vary with similar inputs.
Can generalize to new, unseen data.
Best for problems with complex patterns or variability.
Example: A spam filter learns from labeled emails.
Mistake
AI and machine learning are the same thing.
Correct
Machine learning is a subset of AI. AI encompasses many techniques beyond ML, such as rule-based systems and symbolic reasoning. All ML is AI, but not all AI is ML.
Mistake
AI systems can think and feel like humans.
Correct
Current AI (narrow AI) does not possess consciousness, emotions, or true understanding. It processes data statistically and produces outputs based on patterns. It does not 'think' in a human sense.
Mistake
AI is always objective and unbiased.
Correct
AI models learn from data, which can contain human biases. If training data is biased, the model will reflect those biases. Responsible AI practices aim to mitigate bias, but it is not automatically eliminated.
Mistake
Deep learning is the only type of machine learning.
Correct
Deep learning is a subset of machine learning. Other ML techniques include linear regression, decision trees, random forests, and support vector machines. Deep learning is powerful but not always the best choice.
Mistake
All AI requires massive amounts of data.
Correct
Some AI techniques, like rule-based systems or one-shot learning, can work with little data. However, most modern ML models benefit from large datasets. The amount needed depends on the complexity of the task.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
AI is the overarching field of creating intelligent machines. Machine learning is a subset of AI where systems learn from data without explicit programming. Deep learning is a subset of ML that uses neural networks with many layers to model complex patterns. In short: all deep learning is ML, all ML is AI, but not vice versa. For the exam, know that deep learning is used for tasks like image and speech recognition.
Common workloads include Computer Vision (image analysis, OCR), Natural Language Processing (sentiment, key phrases), Speech (speech-to-text, text-to-speech), Decision (anomaly detection, personalization), and Custom AI (Custom Vision, Custom Speech). Each workload maps to specific Azure AI services. The exam tests your ability to match scenarios to these services.
Supervised learning uses labeled data (input-output pairs) to train a model to predict outputs for new inputs. Examples: classification, regression. Unsupervised learning uses unlabeled data to find hidden patterns or groupings. Examples: clustering, association. On the exam, know that supervised requires labels, unsupervised does not.
The six principles are: Fairness (treat all people fairly), Reliability and Safety (operate reliably), Privacy and Security (protect data), Inclusiveness (empower everyone), Transparency (understand decisions), and Accountability (humans are accountable). These are tested directly; memorize them.
AI models can be biased if trained on biased data. Bias is not automatically removed. Responsible AI practices include auditing data for bias, using diverse datasets, and implementing fairness metrics. The exam may ask you to identify which principle is violated when a model discriminates.
Azure Cognitive Services was renamed to Azure AI Services. They are the same set of pre-built APIs. The exam may use either term. Do not confuse them as different offerings.
Inference is the process of using a trained model to make predictions on new data. It is the deployment phase where the model is used in production. Training is the learning phase; inference is the application phase.
You've just covered What is Artificial Intelligence? — now see how well it sticks with free AI-900 practice questions. Full explanations included, no account needed.
Done with this chapter?